Empirical phenotyping in coupled patient+care systems: Generating low-dimensional categories for hypothesis-driven investigation of mechanically-ventilated patients

Abstract

Background: Analyzing patient data under current mechanical ventilation (MV) management processes is essential to develop hypotheses about improvements and to understand MV consequences over time. However, progress is complicated by the complexity of lung-ventilator system (LVS) interactions, patient-care and patient-ventilator heterogeneity, and a lack of classification schemes for observable behavior.

Method: Ventilator waveform data arise from patient-ventilator interactions within the LVS while care processes manage both patient and ventilator settings. This study develops a computational pipeline that segments these joint waveform data and care settings timeseries to phenotype the data generating process. The modular method supports many methodological choices for representing waveform data and unsupervised clustering.

Results: Applied to 35 ARDS patients including 8 with COVID-19, typcially 8[6.8] (median[IQR]) phenotypes capture 97[3.1]% of data using naive similarity assumptions on waveform and MV settings data. Individual phenotypes organized around ventilator mode, PEEP, and tidal volume with additional segmentation reflecting waveform behaviors. Few (< 10% of) phenotype changes tie to ventilator settings, indicating considerable dynamics in LVS behaviors. Evaluation of phenotype heterogeneity reveals LVS dynamics that cannot be discretized into sub-phenotypes without additional data or alternate assumptions. Suitably normalized individual phenotypes may be aggregated into coherent groupings suitable for analysis of cohort data.

Conclusions: The pipeline is generalizable although empirical output is data- and algorithm-dependent. Further, output phenotypes compactly discretize the data for longitudinal analysis and may be optimized to resolve features of interest for specific applications.

Figure 1:

Broad pipeline organization. Raw data (1.) are digitally parametrized (2.) over short windows, typically satisfying stationarity assumptions. Distributional parameter estimates are summarized and augmented with the contextual data of ventilator settings (3.) which include information such as ventilator operation mode, positive-end expiratory pressure (PEEP) or other baseline pressure, flow and pressure triggers, and minimum mandatory breath rate. Feature vectors, defined by the augmented LVS descriptors, are reduced to three dimensions (4.) where they can be analyzed based on time ordering (top) and structural similarity via segmentation (bottom). Finally, in (5.), temporal evolution of the system is compactly encoded in the time-ordered LVS descriptor labels and their associated waveform characterizations in an interpretable and explainable way. The process transforms raw data (1.) into a more easily comprehensible form such as (5.).

Figure 2:

LVS evolution of patient #34. Panels a–c show the trajectory of phenotype labels, ventilator settings, and externally identified PVD, respectively with a common horizontal axis of patient record hours. In b), non-APVCMV ventilator modes are indicated by shaded regions. The panel (d) shows the model image of waveform parameters nearest to the group median, which characterizes breath pV loops of that phenotype (shown with the same color as (a)). The occurrence of label #1 is discontinuous in time and occurs under different PEEP values suggesting waveform shapes vary only in baseline pressure. The PVD-less evolution of the LVS shows much waveform variation separate from ventilator settings changes. Labeling and coloring in this figure do not relate to other figures.

Figure 3:

A representative example: patient #11. The plot panels (a–d) are the same as the previous figure. A flow trigger (b, at purple arrows) near 3 and 13 hours are the only MV settings changes besides mode, PEEP, or tidal volume. The lower panels (e–g) examine the variability during the record interval 15–21 hours under stationary ventilator settings. The mean (dashed black line) coincides with the golden pV loop (label #10) in the upper plot. The many distinct breath sub-types identified are more similar than to other main types in the upper plot; as a result, they are grouped together at this choice of hyper-parameters. Internal phenotype variability suggests continuous LVS changes that may not admit a natural discretization. Colors coordinate between panels a) and d), and among e), f), and g). Labeling and coloring in this figure do not relate to other figures.

Figure 4:

Membership and data properties associated with cohort phenotypes. Points in panels (a–d) correspond to 721 individual phenotypes shown in UMAP coordinates. Labels (a) mix patients (b) while defining empirical partitions of other factors of patient data (c–h). Groupings separate PEEP (c,g) and ventilator modes (d), which are arguably among the most important ventilator feature elements. Structured distributional separation occurs for continuous breath variables such as tidal volume (e), driving pressure (f), and elastance ($V_T/\left(p_{\mathrm{m}\mathrm{a}\mathrm{x}}-p_{\text{base}}\right)$, g). PEEP (c) and ventilator mode (d) of UMAP labels identify the median value of each individual phenotype; probability densities (e–h) are computed from original data and colored according to panel (a). Labels and colors of panel a) define the those of panels e–h) and Fig 5. Modes: spontaneous (SPONT), Pressure controlled (PC), Synchronized controlled (SC), and Adaptive Pressure Volume Controlled (APVC)

Figure 5:

Non-dimensional waveform characterizations. Pressure-volume traces correspond to median (bold) and nearby (thin) window characterizations of each cohort phenotype. Labels and colors correspond to Fig 4a. Vertical and horizontal scales axes correspond to $\widehat{V}:=V/V_T$ and $\widehat{p}:=\left(p\left(t\right)-p_{\text{base}}\right)/(p_{\text{peak}}-p_{\text{base}})$, respectively, per Fig 4e–g. The dashed line indicates baseline pressure. Cohort phenotypes differentiate waveform shape characteristics and pressure-volume coordination in conjunction with associated scaling factors. Intra-group variation is naturally high given the low specificity of each type.